AI data processing encryption method and device based on deep learning

Abstract

The invention discloses an AI data processing encryption method and device based on deep learning, and relates to the technical field of multi-modal graph embedding, and the method comprises the steps: collecting multi-modal data, and carrying out the preprocessing of the multi-modal data; features of the multi-modal data are extracted, the features of the multi-modal data are fused through a cross attention mechanism, and a multi-modal comprehensive feature vector is generated; constructing an application graph convolutional network GCN relation graph to generate a node embedding vector, and obtaining an enhanced node in combination with the multi-modal comprehensive feature vector; after data encryption is completed, data are protected through secure transmission and distributed storage, and an optimal encryption strategy is selected by using a reinforcement learning model; through a multi-modal graph embedding technology, diversified collection and efficient preprocessing of data are realized, a data relation graph is constructed, node embedding vectors are updated by applying a graph convolutional network GCN, enhanced node embedding vectors are generated in combination with multi-modal comprehensive feature vectors, complex relations among the data are effectively captured, and expressive force of the node embedding vectors is improved.

Description

Technical Field

[0001] The present invention relates to the field of multimodal graph embedding technology, and in particular to a method and device for AI data processing encryption based on deep learning. Background Art

[0002] With the development of big data and artificial intelligence technology, data security and privacy protection have become increasingly important issues; the original data encryption methods mainly rely on fixed key generation mechanisms, such as symmetric encryption and asymmetric encryption based on cryptographic algorithms; however, these methods have certain limitations when facing complex and changing network environments; for example, fixed keys are easy to crack and cannot dynamically adjust to changing threats. In recent years, deep learning technology has demonstrated powerful capabilities in feature extraction, data processing, and pattern recognition, providing new ideas for data encryption; by combining deep learning and graph convolutional networks (GCNs), more efficient data feature fusion and node embedding vector updates can be achieved, thereby generating more complex dynamic key seeds and improving the security of data encryption.

[0003] Existing technologies can usually only process a single type of data and cannot fully utilize information from multiple data sources. Moreover, existing technologies mainly rely on fixed key generation mechanisms, such as symmetric encryption and asymmetric encryption based on cryptographic algorithms. In addition, existing data encryption methods often lack effective processing and feature fusion of multimodal data, making it difficult to capture complex correlations between data. Summary of the invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides an AI data processing encryption method and device based on deep learning to solve the problem of complex correlations between data that are difficult to capture.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In the first aspect, the present invention provides an AI data processing and encryption method based on deep learning, which includes collecting multimodal data and preprocessing the multimodal data; extracting the features of the multimodal data, fusing the multimodal data features through a cross-attention mechanism, and generating a multimodal comprehensive feature vector; constructing a graph convolutional network (GCN) relationship graph to generate a node embedding vector, and combining the multimodal comprehensive feature vector to obtain an enhanced node; generating a complex key seed from the enhanced node, deriving the final key through HKDF, and encrypting the data through an encryption algorithm; after completing the data encryption, protecting the data through secure transmission and distributed storage, and selecting the optimal encryption strategy using a reinforcement learning model.

[0008] As a preferred solution of the AI ​​data processing encryption method based on deep learning described in the present invention, wherein: the multimodal data is collected and preprocessed, and the specific steps are:

[0009] Use web crawlers to collect text data, image data, audio data, and communication records from server logs, user behavior records, and network traffic;

[0010] Perform word segmentation on text data;

[0011] Convert the text after word segmentation to lowercase, remove stop words, and extract stems to obtain clean text data;

[0012] Convert clean text data into vectors;

[0013] All images in the image data are scaled to a uniform size, the image edge information is retained using the linear transformation method, and the images are randomly rotated, flipped, and cropped;

[0014] Use Mel-frequency cepstral coefficients to extract audio features from audio data, use frequency domain filters to reduce noise on audio data, and segment audio signals;

[0015] Cleans up invalid entries and duplicates in communication log data.

[0016] As a preferred solution of the AI ​​data processing encryption method based on deep learning described in the present invention, wherein: the features of multimodal data are extracted, and the features of multimodal data are fused through a cross-attention mechanism to generate a multimodal comprehensive feature vector. The specific steps are as follows:

[0017] Use the BERT model to process the preprocessed text data and generate text feature vectors;

[0018] Use the ResNet-50 model to extract image feature vectors from the preprocessed image data;

[0019] The preprocessed audio data is processed using the GRU model to obtain a hidden state vector;

[0020] The three modal data features are fused through the cross-attention mechanism in multimodal feature fusion to generate a multimodal comprehensive feature vector.

[0021] As a preferred solution of the AI ​​data processing encryption method based on deep learning described in the present invention, wherein: the construction of the application graph convolution network GCN relationship graph to generate node embedding vectors, and combining the multimodal comprehensive feature vector to obtain enhanced nodes, the specific steps are:

[0022] Construct the application graph convolution network GCN data relationship graph based on the preprocessed communication records;

[0023] Use the NetworkX library to create a graph, add nodes and edges, and define node v and edge attributes;

[0024] Through the nonlinear transformation of the activation function σ, the node embedding vector is obtained The expression is:

[0025]

[0026] in, is the embedding vector of node v at layer l, is the embedding vector of neighbor node u at layer l, N(v) is the neighbor set of node v, is the cosine similarity;

[0027] The node embedding vector in the generated application graph convolutional network (GCN) data relationship graph and the multimodal comprehensive feature vector are combined to generate an enhanced node embedding vector.

[0028] As a preferred solution of the AI ​​data processing encryption method based on deep learning described in the present invention, wherein: the enhanced node generates a complex key seed, the final key is derived through HKDF, and the AES-256 encryption algorithm is selected to encrypt the data. The specific steps are:

[0029] The autoregressive AR model introduced into the time series generates a dynamic key seed, which is expressed as:

[0030]

[0031] Among them, S t is the key seed at the current time point, a i is the autoregressive coefficient, S t-i is the past key seed, p is the autoregressive order, i is the lag order index in the time series, ∈ t is the error term;

[0032] Determine the autoregressive order and autoregressive coefficient through historical data training;

[0033] Initialize the past key seed;

[0034] The enhanced node embedding vector is converted into a fixed-length vector and input into the autoregressive AR model to update the key seed at the current time point to obtain the complex dynamic key seed S' t , the expression is:

[0035]

[0036] in, yes The result after hash function processing is is the enhanced node embedding vector;

[0037] The final key is derived from the complex dynamic key seed through the HKDF key derivation function;

[0038] Select AES-256 as the encryption algorithm, set the encryption mode to CBC, generate and use the initialization vector IV;

[0039] Dynamically adjust encryption algorithm parameters based on enhanced node embedding vectors;

[0040] According to the confidentiality level of the data, select the corresponding key and initialization vector for data encryption, and use the AES-256-CBC mode during the encryption process.

[0041] As a preferred solution of the AI ​​data processing encryption method and device based on deep learning described in the present invention, after the data encryption is completed, the data is protected by secure transmission and distributed storage, and the optimal encryption strategy is selected by using a reinforcement learning model. The specific steps are as follows:

[0042] Securely transmit encrypted data via TLS / SSL protocol;

[0043] Use HDFS distributed file storage to store encrypted data in cloud servers and apply redundant coding to improve data reliability;

[0044] The reinforcement learning agent is trained through a reinforcement learning model based on the PPO algorithm, so that the reinforcement learning agent can select the optimal encryption strategy according to the current environment state.

[0045] As a preferred solution of the AI ​​data processing encryption method based on deep learning described in the present invention, wherein: the reinforcement learning agent selects the optimal encryption strategy according to the current environmental state, and the specific steps are:

[0046] Define the state space, action space, and reward function of the encrypted working environment;

[0047] Build a reinforcement learning model based on the PPO algorithm;

[0048] The reinforcement learning model agent selects actions based on the current state of the encrypted working environment through a neural network;

[0049] Apply the selected action to the encryption working environment to encrypt data;

[0050] Give rewards based on the safety score and performance cost in the defined reward function;

[0051] If complex encryption successfully defends against threat attacks, positive rewards will be given;

[0052] If simple encryption leads to performance degradation, negative rewards will be given;

[0053] Set a threshold, and stop training when the reinforcement learning model agent reaches and exceeds the threshold.

[0054] In the second aspect, the present invention provides an AI data processing and encryption device based on deep learning, including a data acquisition module, a data extraction module, a construction module, a key generation module and a transmission optimization module; the data acquisition module is used to collect multimodal data and pre-process the multimodal data; the data extraction module is used to extract the features of the multimodal data, fuse the multimodal data features through a cross-attention mechanism, and generate a multimodal comprehensive feature vector; the construction module is used to construct an application graph convolutional network GCN relationship graph to generate a node embedding vector, and obtain an enhanced node in combination with the multimodal comprehensive feature vector; the key generation module is used to generate a complex key seed from the enhanced node, derive the final key through HKDF, and encrypt the data through an encryption algorithm; the transmission optimization module is used to protect the data through secure transmission and distributed storage after completing the data encryption, and select the optimal encryption strategy using a reinforcement learning model.

[0055] In a third aspect, the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the AI ​​data processing encryption method and apparatus based on deep learning as described in the first aspect of the present invention.

[0056] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the deep learning-based AI data processing and encryption method as described in the first aspect of the present invention.

[0057] The beneficial effects of the present invention are as follows: through multimodal graph embedding technology, diversified collection and efficient preprocessing of data are realized, and the quality and consistency of data are improved; a data relationship graph is constructed and the graph convolutional network GCN is applied to update the node embedding vector, and the multimodal comprehensive feature vector is combined to generate an enhanced node embedding vector, which effectively captures the complex relationship between data and improves the expressiveness of the node embedding vector; then, a complex dynamic key seed is generated through the autoregressive AR model, and the final key is derived through HKDF, and the AES-256 encryption algorithm is selected to encrypt the data, which significantly improves the randomness and unpredictability of the key and enhances the security of data encryption. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0059] Figure 1 This is a flow chart of the AI ​​data processing and encryption method based on deep learning in Example 1.

[0060] Figure 2 This is a flow chart of the distributed storage and optimized encrypted data strategy in Example 1. DETAILED DESCRIPTION

[0061] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0062] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0063] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0064] Example 1, reference Figure 1 and Figure 2 , which is the first embodiment of the present invention, and provides an AI data processing encryption method based on deep learning, comprising the following steps:

[0065] S1. Collect multimodal data and preprocess the multimodal data.

[0066] Going a step further, we use web crawlers to collect text data, image data, audio data, and communication records from server logs, user behavior records, and network traffic;

[0067] For Chinese text in the text data, the jieba library is used for word segmentation, and for English text, the nltk library is used for word segmentation;

[0068] It should be noted that word segmentation is the basis of natural language processing, which divides continuous text into meaningful vocabulary units to facilitate subsequent text processing and analysis;

[0069] Convert the text data after word segmentation into lowercase, remove stop words, and extract stems to obtain sense text data;

[0070] It should be noted that the segmented text was uniformly converted to lowercase, common stop words were removed, and stemming techniques were used to extract word roots;

[0071] Use the Word2Vec model to convert clean text data into vectors;

[0072] It should be noted that the Word2Vec model is used to convert the segmented text data into a high-dimensional vector representation. Word2Vec is a commonly used word embedding technology that can map words to a high-dimensional vector space, in which semantically similar words are closer in the vector space;

[0073] All images in the image data are scaled to a uniform size, the image edge information is retained using the linear transformation method, and the images are randomly rotated, flipped, and cropped;

[0074] The audio data is extracted using Mel-frequency cepstral coefficients, the audio data is denoised using frequency domain filters, and the audio signal is segmented into multiple segments;

[0075] It should be noted that long audio signals are broken down into smaller segments that are easier to process;

[0076] Cleans up invalid entries and duplicates in communication log data.

[0077] It should be noted that, through the multimodal data processing method of the present invention, efficient, comprehensive and high-quality data collection and preprocessing are achieved; web crawlers are used to automatically collect text, images, audio and communication records from multiple data sources, which improves the efficiency and coverage of data collection and ensures the diversity and comprehensiveness of data; text data is segmented, converted to lowercase, stop words are removed and stems are extracted, the text data is standardized, noise is reduced, and the accuracy and efficiency of feature extraction are improved; the Word2Vec model is used to convert clean text data into vectors, semantic information is retained, and the performance of text analysis is enhanced; image data is resized, edge information is retained and randomly transformed, which increases data diversity and improves the generalization ability and robustness of the model; MFCC features are extracted from audio data and noise reduction and segmentation are performed, which effectively captures acoustic characteristics and improves the quality of audio processing; invalid entries and duplicates in communication records are cleaned up, and data consistency and reliability are improved.

[0078] S2. Extract the features of multimodal data, fuse the features of multimodal data through the cross-attention mechanism, and generate a multimodal comprehensive feature vector.

[0079] Furthermore, the preprocessed text data is processed using the BERT model to generate a 768-dimensional text feature vector;

[0080] It should be noted that BERT is a pre-trained language model based on the Transformer architecture, which can capture contextual information and semantic relationships in text;

[0081] Each 768-dimensional vector represents the characteristics of a text segment (such as a sentence or paragraph); these vectors are obtained by encoding the text, and each dimension represents the characteristics of the text in a specific aspect;

[0082] The ResNet-50 model is used to extract the 2048-dimensional feature vector of the image from the preprocessed image data;

[0083] It should be noted that ResNet-50 is a deep convolutional neural network with multiple convolutional layers and residual connections, which can capture high-level abstract features in images;

[0084] Each 2048-dimensional vector represents the characteristics of an image; these vectors are obtained from the output of the last fully connected layer of the image, and each dimension represents the characteristics of the image in a specific aspect.

[0085] The preprocessed audio data is processed using the GRU model to obtain a 128-dimensional hidden state vector;

[0086] It should be noted that GRU is a recurrent neural network that is particularly suitable for processing sequence data, such as audio signals;

[0087] Each 128-dimensional vector represents the hidden state of the audio data. These vectors are obtained from the last time step output of the GRU model, and each dimension represents the characteristics of the audio in a specific aspect.

[0088] The three modal data features are fused through the cross-attention mechanism in multimodal feature fusion to generate a multimodal comprehensive feature vector M.

[0089] It should be noted that high-quality feature extraction and effective fusion are achieved through multimodal data processing methods; the BERT model is used to generate a 768-dimensional text feature vector, which retains rich semantic information and improves the quality and expressiveness of text features; the ResNet-50 model is used to extract a 2048-dimensional image feature vector, which comprehensively captures the key features of the image and enhances the richness and accuracy of the image features; the GRU model is used to process audio data to generate a 128-dimensional hidden state vector, which effectively captures the temporal dynamic characteristics of the audio and improves the representativeness of the audio features; the data features of the three modalities are fused through the cross-attention mechanism to generate a multimodal comprehensive feature vector M, which not only combines the advantages of each modality, but also captures the correlation and complementarity between them; ultimately, the effective processing of multimodal data and the generation of high-quality features are jointly ensured, which significantly improves the performance and accuracy of the system and provides a solid foundation for subsequent data encryption and secure transmission.

[0090] S2.1. The three modal data features are fused through the cross attention mechanism in multimodal feature fusion to generate a multimodal comprehensive feature vector M, which is expressed as:

[0091] M = Attn(T, I, A);

[0092] Among them, T represents the text feature vector, I represents the image feature vector, and A represents the audio feature vector.

[0093] S3. Build and apply the graph convolutional network (GCN) relationship graph to generate node embedding vectors, and combine the multimodal comprehensive feature vectors to obtain enhanced nodes.

[0094] Furthermore, the application graph convolutional network GCN data relationship graph is constructed based on the preprocessed communication records;

[0095] Use the NetworkX library to create a graph, add nodes and edges, and define node v and edge attributes;

[0096] Initialize an embedding vector for each node in the GCN data relationship graph

[0097] Calculate the cosine similarity between each node in the GCN data relationship graph and all neighbor nodes u, expressed as:

[0098]

[0099] in, is the cosine similarity, is the embedding vector of node v at layer l, is the embedding vector of neighbor node u at layer l;

[0100] Using cosine similarity as weight, embed the neighbor nodes into vector Perform weighting to obtain the weighted sum vector, expressed as:

[0101]

[0102] Where N(v) is the neighbor set of node v;

[0103] The weighted sum vector is transformed nonlinearly through the activation function σ to obtain a new node embedding vector The expression is:

[0104]

[0105] Combine the new node embedding vector and the multimodal comprehensive feature vector in the generated application graph convolutional network GCN data relationship graph to generate an enhanced node embedding vector The expression is:

[0106]

[0107] It should be noted that by constructing and applying the graph convolutional network (GCN) data relationship graph, the entities in the communication records and their complex associations are structured and represented, which enhances the interpretability and processing capabilities of the data; initializing the embedding vector for each node provides the basis for the graph convolution operation, enabling the system to gradually optimize the node features; the cosine similarity is used to quantify the relationship strength between nodes and neighbors, which effectively improves the quality and accuracy of node embedding; finally, by fusing multimodal comprehensive feature vectors to generate enhanced node embedding, not only the graph structure information is retained, but also the rich features from different data sources are integrated, significantly improving the expressiveness and discrimination of node representation.

[0108] S3.1. The node attributes include IP address, geographic location and device type;

[0109] Edge attributes include communication times, communication time, and communication protocol type.

[0110] IP address of node attribute: One attribute of a node is its corresponding IP address;

[0111] Geographic location of node attributes: Another attribute of a node is its geographic location information;

[0112] Device type of node attributes: The third attribute of a node is the device type (e.g., server, mobile device, etc.);

[0113] Communication times of edge attributes: One attribute of an edge is the number of communications between two nodes;

[0114] Communication time of edge attributes: Another attribute of an edge is the timestamp of the most recent communication;

[0115] Communication protocol type of edge attribute: The third attribute of the edge is the type of communication protocol used (such as TCP, UDP, etc.).

[0116] S4. Generate a complex key seed by enhancing the node, derive the final key through HKDF, and encrypt the data through the encryption algorithm.

[0117] Furthermore, the autoregressive AR model in the time series is introduced to generate a dynamic key seed, which is expressed as:

[0118]

[0119] Among them, S t is the key seed at the current time point, a i is the autoregressive coefficient, S t-i is the past key seed, p is the autoregressive order, i is the lag order index in the time series, ∈ t is the error term;

[0120] Determine the autoregressive order and autoregressive coefficient through historical data training;

[0121] Initialize the past key seed;

[0122] The enhanced node embedding vector is converted into a fixed-length vector and input into the autoregressive AR model to update the key seed at the current time point to obtain the complex dynamic key seed S' t , the expression is:

[0123]

[0124] in, yes The result after being processed by the hash function;

[0125] The final key is derived from the complex dynamic key seed by the HKDF key derivation function, and the expression is:

[0126] K=HKDF(S' t ,L);

[0127] Where K is the final key and L is the output key length;

[0128] Select AES-256 as the encryption algorithm, set the encryption mode to CBC, generate and use the initialization vector IV;

[0129] Dynamically adjust encryption algorithm parameters based on enhanced node embedding vectors;

[0130] According to the confidentiality level of the data, select the corresponding key and initialization vector for data encryption, and use the AES-256-CBC mode during the encryption process to ensure data security.

[0131] It should be noted that by introducing the autoregressive AR model to generate dynamic key seeds, the time dependence and dynamics of the key are increased, and the security and anti-prediction ability of the key are improved; the enhanced node embedding vector is added to the key seed generation process, which further increases the complexity and diversity of the key and improves the security of the key; HKDF is used to derive the final key from the complex dynamic key seed, which ensures the high quality and randomness of the key and provides additional security; AES-256 is selected as the encryption algorithm, and the encryption mode is set to CBC, and the initialization vector IV is generated and used to ensure the high security and integrity of the data during the encryption process and prevent the data from being tampered with or leaked.

[0132] S4.1. According to Dynamically adjust the key length based on certain characteristics of the key (such as norm, value of a specific dimension, etc.);

[0133] like The norm of is larger, and longer keys can be used to improve security;

[0134] according to For some features of , dynamically select or generate different initialization vectors; for example, Generate IV through hash function;

[0135] according to Select different encryption modes (such as CBC, CTR, etc.) based on certain characteristics of If the value of some dimension is high, a more secure encryption mode is selected.

[0136] S4.2. It should be noted that the autoregressive AR model is used to generate a complex dynamic key seed using an enhanced node embedding vector, and the final key K is derived in combination with the HKDF algorithm to ensure the high security and randomness of the key; subsequently, the AES-256 encryption algorithm is used to encrypt the data in CBC mode with the initialization vector IV to achieve data security protection; this process not only enhances the unpredictability of the key generation mechanism, but also adapts to different confidentiality requirements by dynamically adjusting encryption parameters, thereby providing a powerful and flexible data encryption solution.

[0137] S5. After completing data encryption, protect the data through secure transmission and distributed storage, and use the reinforcement learning model to select the optimal encryption strategy.

[0138] Furthermore, encrypted data is transmitted securely through TLS / SSL protocols, and digital signatures and certificates are used to ensure data integrity and source authenticity;

[0139] Use HDFS distributed file storage to store encrypted data in cloud servers and apply redundant coding to improve data reliability;

[0140] The reinforcement learning agent is trained through a reinforcement learning model based on the PPO algorithm, so that the reinforcement learning agent can select the optimal encryption strategy according to the current environment state;

[0141] It should be noted that the reinforcement learning agent dynamically adjusts the encryption algorithm and parameters to adapt to environmental changes and improve overall security and performance.

[0142] It should be noted that the combined use of TLS / SSL protocols and digital signatures and certificates ensures the confidentiality, integrity and source credibility of data during transmission, significantly improving the security of data transmission; using HDFS distributed file storage and redundant coding technology, encrypted data is stored in cloud servers in a dispersed manner, enhancing data storage reliability and fault tolerance, and ensuring high data availability and recovery capabilities; training the intelligent agent through a reinforcement learning model based on the PPO algorithm enables the intelligent agent to dynamically select the optimal encryption strategy based on the current environmental state, thereby improving adaptability and robustness; reinforcement learning agents can adjust encryption algorithms and parameters in real time to adapt to changing security threats and performance requirements, which not only improves the security of the system, but also optimizes overall performance.

[0143] S5.1. The reinforcement learning agent selects the optimal encryption strategy according to the current environment state. The specific steps are:

[0144] Define the state space, action space, and reward function of the encrypted working environment;

[0145] Build a reinforcement learning model based on the PPO algorithm;

[0146] The reinforcement learning model agent selects actions based on the current state of the encrypted working environment through a neural network;

[0147] Apply the selected action to the encryption working environment to encrypt data;

[0148] Give rewards based on the safety score and performance cost in the defined reward function;

[0149] If complex encryption successfully defends against threat attacks, positive rewards will be given;

[0150] If simple encryption leads to performance degradation, negative rewards will be given;

[0151] Set a threshold, and stop training when the reinforcement learning model agent reaches and exceeds the threshold.

[0152] S5.2. It should be noted that HDFS is mainly used for big data processing, especially in the fields of batch data processing and data analysis; it is usually used in conjunction with distributed computing frameworks such as MapReduce and Apache Spark to handle tasks such as log analysis, data warehouses, machine learning, and large-scale data processing.

[0153] S5.3. Use the PPO algorithm to train the reinforcement learning agent. The specific steps are as follows:

[0154] Define state space, action space, and reward function;

[0155] State space: The state of the environment includes network traffic characteristics, user behavior characteristics, system resource usage, etc. This state information can help the agent understand the current security threat level and performance requirements;

[0156] Action space: The action space includes different encryption algorithms (such as AES-256), encryption modes (such as CBC), key lengths, and initialization vector selections. The agent needs to select an action from this action space to execute;

[0157] Reward function: Design a reward function to evaluate the agent's actions; the reward function usually takes into account the security score (e.g., the ability to resist attacks) and performance overhead (e.g., the speed of encryption / decryption). The goal is to find a balance that ensures data security without affecting the performance of the system;

[0158] Build a reinforcement learning model based on the PPO algorithm;

[0159] Choose an agent as the model architecture. The agent is a neural network that takes the state of the environment as input and outputs a probability distribution over actions. For discrete action spaces, we can use a network with categorical outputs.

[0160] The PPO algorithm is used to adjust the strategy. PPO is a policy gradient method that maximizes the cumulative reward by optimizing the strategy. PPO introduces a clipping objective function to limit the step size of each update, thereby improving the stability and efficiency of training. The PPO algorithm calculates the advantage function at each time step and adjusts the strategy based on it.

[0161] The agent interacts with the environment, generating a sequence of state-action-reward-new state;

[0162] Calculate the advantage function for each action based on the reward and value function estimates;

[0163] Use the objective function of PPO to update the parameters of the policy network; PPO gradually improves the policy through multiple small batch updates;

[0164] Repeat the above process until the agent learns to choose appropriate encryption strategies in different environments;

[0165] After training is completed, the agent can select the optimal encryption strategy in real time according to the current environment status during actual operation;

[0166] The agent selects the most appropriate encryption algorithm, key length, and initialization vector based on the current environment status (such as network traffic characteristics, user behavior characteristics, etc.). For example, when a high-risk attack is detected, the agent may choose a more complex encryption algorithm and a longer key length; while in low-risk and high-performance situations, the agent may choose a simpler encryption scheme;

[0167] The agent can also continue to learn from new interactions, constantly adjusting and optimizing its strategy to adapt to changing environmental conditions;

[0168] Deploy the trained agent to production, ensuring it has access to the required state information and can perform the corresponding cryptographic operations;

[0169] Continuously monitor the agent’s performance, including safety and performance metrics. If the agent is found to be underperforming, it may be necessary to retrain or adjust the reward function.

[0170] The present embodiment also provides an AI data processing and encryption device based on deep learning, including: a data acquisition module, a data extraction module, a construction module, a key generation module and a transmission optimization module; the data acquisition module is used to collect multimodal data and pre-process the multimodal data; the data extraction module is used to extract the features of the multimodal data, fuse the multimodal data features through a cross-attention mechanism, and generate a multimodal comprehensive feature vector; the construction module is used to construct an application graph convolutional network GCN relationship graph to generate a node embedding vector, and obtain an enhanced node in combination with the multimodal comprehensive feature vector; the key generation module is used to generate a complex key seed from the enhanced node, derive the final key through HKDF, and encrypt the data through an encryption algorithm; the transmission optimization module is used to protect the data through secure transmission and distributed storage after completing the data encryption, and select the optimal encryption strategy using a reinforcement learning model.

[0171] This embodiment also provides a computer device, which is suitable for the AI ​​data processing and encryption method based on deep learning, including: a memory and a processor; the memory is used to store computer executable instructions, and the processor is used to execute computer executable instructions to implement the AI ​​data processing and encryption method based on deep learning proposed in the above embodiment.

[0172] The computer device may be a terminal, and the computer device includes a processor, a memory, a communication interface, a display screen and an input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. The display screen of the computer device may be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device may be a touch layer covering the display screen, or a key, trackball or touchpad provided on the housing of the computer device, or an external keyboard, touchpad or mouse, etc.

[0173] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by the processor, the AI ​​data processing and encryption method based on deep learning proposed in the above embodiment is implemented; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (Static Random Access Memory, referred to as SRAM), electrically erasable programmable read-only memory (Electrically Erasable Programmable Read-Only Memory, referred to as EEPROM), erasable programmable read-only memory (Erasable Programmable Read Only Memory, referred to as EPROM), programmable read-only memory (Programmable Red-Only Memory, referred to as PROM), read-only memory (Read-Only Memory, referred to as ROM), magnetic storage, flash memory, disk or optical disk.

[0174] In summary, the present invention realizes diversified collection and efficient preprocessing of data through multimodal graph embedding technology, thereby improving the quality and consistency of data; constructs a data relationship graph and applies graph convolutional network GCN to update node embedding vectors, and combines multimodal comprehensive feature vectors to generate enhanced node embedding vectors, thereby effectively capturing the complex relationship between data and improving the expressiveness of node embedding vectors; then, a complex dynamic key seed is generated through the autoregressive AR model, and the final key is derived through HKDF, and the AES-256 encryption algorithm is selected to encrypt the data, which significantly improves the randomness and unpredictability of the key and enhances the security of data encryption.

[0175] Example 2, referring to Table 1, is the second example of the present invention. In order to further verify the technical solution of the present invention, experimental simulation data of the AI ​​data processing encryption method based on deep learning is provided.

[0176] First, we use web crawlers to collect multimodal data such as text, images, audio, and communication records from server logs, user behavior records, and network traffic, and perform preprocessing. This includes: using the jieba library to segment Chinese text and the nltk library to segment English text; converting text to lowercase, removing stop words, and extracting stems; using the Word2Vec model to convert text data into vectors; uniformly scaling image data to 224x224 pixels, using the linear transformation method to retain edge information, and performing random rotation, flipping, and cropping; using the Mel-frequency cepstral coefficients (MFCC) to extract features from audio data, and using frequency domain filters to reduce noise;

[0177] Secondly, the BERT model is used to process text data to generate a 768-dimensional text feature vector, the ResNet-50 model is used to extract a 2048-dimensional image feature vector, and the GRU model is used to process audio data to obtain a 128-dimensional hidden state vector. The data features of the three modalities are fused through the cross-attention mechanism to generate a multimodal comprehensive feature vector.

[0178] Next, the NetworkX library is used to create a graph structure based on the preprocessed communication records, define node and edge attributes, and initialize the embedding vector of each node. The cosine similarity between the node and its neighboring nodes is calculated, and a weighted sum is performed based on this. The new node embedding vector is obtained through nonlinear transformation of the activation function, and then combined with the multimodal comprehensive feature vector to generate an enhanced node embedding vector.

[0179] Finally, a complex dynamic key seed is generated through the autoregressive AR model, and the final key is derived in combination with the HKDF algorithm; AES-256 is selected as the encryption algorithm, and the data is encrypted in CBC mode with the initialization vector IV; the encrypted data is securely transmitted through the TLS / SSL protocol, and the encrypted data is stored in the cloud using HDFS, and redundant coding is applied to improve data reliability; the reinforcement learning model based on the PPO algorithm is used to train the reinforcement learning agent, so that it can dynamically adjust the encryption algorithm and parameters according to the current environmental status, thereby improving the overall security and performance.

[0180] The details are shown in Table 1 below:

[0181] Table 1 Data processing and encryption analysis table

[0182]

[0183]

[0184] Significant improvements have been shown in multimodal data processing and encryption technology. In terms of comprehensive feature vector generation, the present invention adopts a cross-attention mechanism to fuse multimodal features. Compared with the simple splicing method in the prior art, the efficiency is improved from 7.5 to 8.9, an increase of 18.7%. In terms of node embedding vector update, the present invention replaces the traditional static node embedding vector with GCN combined with multimodal comprehensive feature vectors, and the efficiency is improved from 6.8 to 8.4, an increase of 23.5%. In terms of key generation, the present invention uses an autoregressive AR model and HKDF algorithm to replace the fixed key length and static encryption parameters, and the efficiency is improved from 7.2 to 9.1, an increase of 26.4%. In terms of image feature extraction, the present invention adopts the ResNet-50 model. Compared with the traditional convolutional neural network CNN, the accuracy is improved from 85.2% to 90.7%, an increase of 6.2%.

[0185] These technologies not only improve processing efficiency and accuracy in all aspects, but also enhance the overall performance and security of the system, providing a more efficient and secure solution for data processing and encryption.

[0186] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. An AI data processing encryption method based on deep learning, characterized in that: include, Collect multimodal data and preprocess the multimodal data; Extract the features of multimodal data, fuse the features of multimodal data through the cross-attention mechanism, and generate a multimodal comprehensive feature vector; Construct and apply the graph convolutional network (GCN) relationship graph to generate node embedding vectors, and combine the multi-modal comprehensive feature vectors to obtain enhanced nodes; The enhanced node generates a complex key seed, the final key is derived through HKDF, and the data is encrypted through the encryption algorithm; After data encryption is completed, the data is protected through secure transmission and distributed storage, and the reinforcement learning model is used to select the optimal encryption strategy.

2. The AI ​​data processing and encryption method based on deep learning as claimed in claim 1, characterized in that: The multimodal data is collected and preprocessed, and the specific steps are: Use web crawlers to collect text data, image data, audio data, and communication records from server logs, user behavior records, and network traffic; Perform word segmentation on text data; Convert the text after word segmentation to lowercase, remove stop words, and extract stems to obtain clean text data; Convert clean text data into vectors; All images in the image data are scaled to a uniform size, the image edge information is retained using the linear transformation method, and the images are randomly rotated, flipped, and cropped; Use Mel-frequency cepstral coefficients to extract audio features from audio data, use frequency domain filters to reduce noise on audio data, and segment audio signals; Cleans up invalid entries and duplicates in communication log data.

3. The AI ​​data processing and encryption method based on deep learning as claimed in claim 2, characterized in that: The method of extracting the features of multimodal data and fusing the features of multimodal data through a cross-attention mechanism to generate a multimodal comprehensive feature vector comprises the following specific steps: Use the BERT model to process the preprocessed text data and generate text feature vectors; Use the ResNet-50 model to extract image feature vectors from the preprocessed image data; The preprocessed audio data is processed using the GRU model to obtain a hidden state vector; The three modal data features are fused through the cross-attention mechanism in multimodal feature fusion to generate a multimodal comprehensive feature vector.

4. The AI ​​data processing and encryption method based on deep learning as claimed in claim 3, characterized in that: The construction and application of the graph convolutional network (GCN) relationship graph generates a node embedding vector, and combines the multi-modal comprehensive feature vector to obtain an enhanced node. The specific steps are: Construct the application graph convolution network GCN data relationship graph based on the preprocessed communication records; Use the NetworkX library to create a graph, add nodes and edges, and define node v and edge attributes; Through the nonlinear transformation of the activation function σ, the node embedding vector is obtained The expression is: in, is the embedding vector of node v at layer l, is the embedding vector of neighbor node u at layer l, N(v) is the neighbor set of node v, is the cosine similarity; The node embedding vector in the generated application graph convolutional network (GCN) data relationship graph and the multimodal comprehensive feature vector are combined to generate an enhanced node embedding vector.

5. The AI ​​data processing and encryption method based on deep learning as claimed in claim 4, characterized in that: The enhanced node generates a complex key seed, the final key is derived through HKDF, and the AES-256 encryption algorithm is selected to encrypt the data. The specific steps are: The autoregressive AR model introduced into the time series generates a dynamic key seed, which is expressed as: Among them, S t is the key seed at the current time point, a i is the autoregressive coefficient, S t-i is the past key seed, p is the autoregressive order, i is the lag order index in the time series, ∈ t is the error term; Determine the autoregressive order and autoregressive coefficient through historical data training; Initialize the past key seed; The enhanced node embedding vector is converted into a fixed-length vector and input into the autoregressive AR model to update the key seed at the current time point to obtain the complex dynamic key seed S' t , the expression is: in, yes The result after hash function processing is is the enhanced node embedding vector; The final key is derived from the complex dynamic key seed through the HKDF key derivation function; Select AES-256 as the encryption algorithm, set the encryption mode to CBC, generate and use the initialization vector IV; Dynamically adjust encryption algorithm parameters based on enhanced node embedding vectors; According to the confidentiality level of the data, select the corresponding key and initialization vector for data encryption, and use the AES-256-CBC mode during the encryption process.

6. The AI ​​data processing and encryption method based on deep learning as claimed in claim 5, characterized in that: After completing the data encryption, the data is protected through secure transmission and distributed storage, and the optimal encryption strategy is selected using the reinforcement learning model. The specific steps are as follows: Securely transmit encrypted data via TLS / SSL protocol; Use HDFS distributed file storage to store encrypted data in cloud servers and apply redundant coding to improve data reliability; The reinforcement learning agent is trained through a reinforcement learning model based on the PPO algorithm, so that the reinforcement learning agent can select the optimal encryption strategy according to the current environment state.

7. The AI ​​data processing and encryption method based on deep learning as claimed in claim 6, characterized in that: The reinforcement learning agent selects the optimal encryption strategy according to the current environment state. The specific steps are: Define the state space, action space, and reward function of the encrypted working environment; Build a reinforcement learning model based on the PPO algorithm; The reinforcement learning model agent selects actions based on the current state of the encrypted working environment through a neural network; Apply the selected action to the encryption working environment to encrypt data; Give rewards based on the safety score and performance cost in the defined reward function; If complex encryption successfully defends against threat attacks, positive rewards will be given; If simple encryption leads to performance degradation, negative rewards will be given; Set a threshold, and stop training when the reinforcement learning model agent reaches and exceeds the threshold.

8. An AI data processing and encryption device based on deep learning, based on the AI ​​data processing and encryption method based on deep learning according to any one of claims 1 to 7, characterized in that: Including, data collection module, data extraction module, construction module, key generation module and transmission optimization module; The data acquisition module is used to collect multimodal data and preprocess the multimodal data; The data extraction module is used to extract the features of multimodal data, fuse the features of multimodal data through a cross-attention mechanism, and generate a multimodal comprehensive feature vector; The construction module is used to construct a graph convolutional network (GCN) relationship graph to generate a node embedding vector, and combine the multimodal comprehensive feature vector to obtain an enhanced node; The key generation module is used to generate a complex key seed from the enhanced node, derive the final key through HKDF, and encrypt the data through the encryption algorithm; The transmission optimization module is used to protect data through secure transmission and distributed storage after completing data encryption, and to select the optimal encryption strategy using a reinforcement learning model.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the deep learning-based AI data processing encryption method described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the deep learning-based AI data processing encryption method according to any one of claims 1 to 7 are implemented.

Images